Optimal DR-Submodular Maximization and Applications to Provable Mean Field Inference